Qwen · 2026-08 · Dense · Multimodal
Qwen3.8 27B
Qwen3.8 27B is a grouped-query attention (GQA) transformer released by Qwen in 2026-08, with 64 layers, hidden size 5120 and a context window of 262,144 tokens.
Layer stack
Linear ×48 Attention ×16 · Layers 64
Key facts
| Family | Qwen |
|---|---|
| Released | 2026-08 |
| Params | 23.7 B |
| Context | 262,144 tokens |
| Attention | GQA (24:4) |
| Layers | 64 |
| Hidden | 5,120 |
| Heads | 24 |
| Vocab | 248,320 |
| Position | RoPE |
| Norm | RMSNorm |
| Activation | silu |
| Dtype | bf16 |
| Architecture class | Qwen3_5ForConditionalGeneration |
Architecture overview
Attention
grouped-query attention (GQA) — 24 q-heads / 4 kv-heads.
Feed-forward / MoE
It is a dense model: all 23.7B parameters participate in computing every token.
The configuration ships 1 extra multi-token-prediction (MTP) layers used to accelerate decoding.
Field-level comparison
Compared against the previous model of the same lab; where none exists, against the structurally closest model. The ratio column is this model divided by the comparison model.
Full-field comparison vs predecessor Qwen3.8 2.4T A95B
| Model | Qwen3.8 27B | Qwen3.8 2.4T A95B | Ratio |
|---|---|---|---|
| model_type | qwen3_5 | qwen3_5_moe_text | ≠ |
| architectures | Qwen3_5ForConditionalGeneration | Qwen3_5MoeForCausalLM | ≠ |
| hidden_size | 5120 | 8192 | ×0.625 |
| num_hidden_layers | 64 | 92 | ×0.696 |
| num_attention_heads | 24 | 64 | ×0.375 |
| num_key_value_heads | 4 | 4 | ≈1 |
| head_dim | 256 | 256 | ≈1 |
| intermediate_size | 17408 | — | |
| hidden_act | silu | silu | ＝ |
| num_experts_per_tok | — | 10 | |
| moe_intermediate_size | — | 2048 | |
| max_position_embeddings | 262144 | 262144 | ≈1 |
| rms_norm_eps | 0.000001 | 0.000001 | ≈1 |
| vocab_size | 248320 | 248320 | ≈1 |
| image_token_id | 248056 | — | |
| language_model_only | false | — | |
| tie_word_embeddings | false | false | ＝ |
| transformers_version | 5.8.0.dev0 | 4.57.3 | ≠ |
| video_token_id | 248057 | — | |
| vision_end_token_id | 248054 | — | |
| vision_start_token_id | 248053 | — | |
| attention_bias | false | false | ＝ |
| attention_dropout | 0 | 0 | ＝ |
| attn_output_gate | true | true | ＝ |
| bos_token_id | 248044 | 248044 | ≈1 |
| dtype | bfloat16 | bfloat16 | ＝ |
Most similar architectures
- Qwen3.6 27B (Qwen, 2026-04) 1.000
- Qwen3 32B (Qwen, 2025-04) 0.998
- Llama 3.1 70B (Meta, 2024-07) 0.994
- Qwen3 8B (Qwen, 2025-04) 0.992
- Llama 3.1 8B (Meta, 2024-07) 0.992
- Hunyuan-TurboS 7B (Hunyuan, 2026-05) 0.992
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
39 fields
| architectures | Qwen3_5ForConditionalGeneration |
|---|---|
| image_token_id | 248056 |
| language_model_only | false |
| model_type | qwen3_5 |
| tie_word_embeddings | false |
| transformers_version | 5.8.0.dev0 |
| video_token_id | 248057 |
| vision_end_token_id | 248054 |
| vision_start_token_id | 248053 |
| attention_bias | false |
| attention_dropout | 0 |
| attn_output_gate | true |
| bos_token_id | 248044 |
| dtype | bfloat16 |
| eos_token_id | 248044 |
| full_attention_interval | 4 |
| head_dim | 256 |
| hidden_act | silu |
| hidden_size | 5120 |
| initializer_range | 0.02 |
| intermediate_size | 17408 |
| layer_types | linear_attention×48 + full_attention×16 |
| linear_conv_kernel_dim | 4 |
| linear_key_head_dim | 128 |
| linear_num_key_heads | 16 |
| linear_num_value_heads | 48 |
| linear_value_head_dim | 128 |
| mamba_ssm_dtype | float32 |
| max_position_embeddings | 262144 |
| mtp_num_hidden_layers | 1 |
| mtp_use_dedicated_embeddings | false |
| num_attention_heads | 24 |
| num_hidden_layers | 64 |
| num_key_value_heads | 4 |
| output_gate_type | swish |
| partial_rotary_factor | 0.25 |
| rms_norm_eps | 0.000001 |
| use_cache | true |
| vocab_size | 248320 |